https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69983
--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> ---
comment #2 (In reply to Richard Biener from comment #2)
> So we fail analyzing the loop nest because a) the loop header check of the
> inner
> loop makes it conditionally executed, b) we have outer loop IV computes i+-1
> guarded by that check.
>
> For a/b) it would be best if the outer loop were tail duplicated.
>
> Looking at the other reported fallout now.
That's unrelated.
So the CFG is the following.
<bb 7>:
<bb 8>:
# i_34 = PHI <1(3), prephitmp_55(7)>
if (_49 > 1)
goto <bb 10>;
else
goto <bb 9>;
<bb 10>:
_52 = i_34 + -1;
_53 = i_34 + 1;
goto <bb 5>;
<bb 4>:
<bb 5>:
# j_35 = PHI <1(10), _14(4)>
_10 = P[i_34][j_35];
_11 = j_35 + -1;
_12 = P[i_34][_11];
_13 = _10 + _12;
_14 = j_35 + 1;
_15 = P[i_34][_14];
_16 = _13 + _15;
_18 = P[_52][j_35];
_19 = _16 + _18;
_21 = P[_53][j_35];
_22 = _19 + _21;
_23 = _22 / 5.0e+0;
P[i_34][j_35] = _23;
if (_14 < _49)
goto <bb 4>;
else
goto <bb 6>;
<bb 9>:
_54 = i_34 + 1;
goto <bb 6>;
<bb 6>:
# prephitmp_55 = PHI <_53(5), _54(9)>
if (_30 > prephitmp_55)
goto <bb 7>;
else
goto <bb 11>;
<bb 7>:
<bb 11>:
return;
and hoisting the loop header check out of the outer loop fixes the PR
(thus for example run it at -O3). loop unswitching performs this
optimization. Not sure if we can improve things otherwise - the IV
computations of i + 1 / i - 1 are inside the inner loop and thus guarded
by the check. The only thing we can do is try to improve the overflow
detection in SCEV / niter analysis to conclude the now unsigned IV { 1, +, 1
}_1
does not overflow because of the loop header check:
_30 = N1_6(D) + -1;
if (_30 > 1)
goto <bb 3>;
else
goto <bb 11>;
that would be done via chrec_convert / convert_affine_scev here. Sadly
the outer loop has no control-IVs recorded for example because of the
redundant _53 / _54 IV update fed into the PHI in bb 6. Looks like that
mess is caused by PRE first hoisting the loop invariant i + 1 / i - 1
and then PREing it (but not hoisting because it doesn't implement that...).
